Upload folder using huggingface_hub

README.md

@@ -8,131 +8,130 @@ language:
 - nl
 - it
 tags:
-- multilingual
-- text-classification
-- content-filtering
-- multi-dimensional-scoring
-- knowledge-distillation
-- uplifting-content
-- news-analysis
+- base_model:adapter:Qwen/Qwen2.5-1.5B
+- lora
+- transformers
 library_name: peft
 base_model: Qwen/Qwen2.5-1.5B
 pipeline_tag: text-classification
 ---
-
-# Uplifting Content Filter v5
-
-## Model Description
-
-A fine-tuned **Qwen2.5-1.5B** model with LoRA adapters for multi-dimensional uplifting content scoring.
-
-This model evaluates news articles across **6 orthogonal dimensions** to identify genuinely uplifting content with documented positive outcomes - not just feel-good stories or speculation.
-
-**Key Innovation**: Uses an orthogonal dimension framework (inspired by LCSA methodology) to avoid the high correlation issues found in previous versions.
-
-## Dimensions
-
-The model scores articles on 6 dimensions:
-
-### Impact Domains (WHAT kind of uplift)
-| Dimension | Weight | Question |
-|-----------|--------|----------|
-| **Human Wellbeing Impact** | 25% | Health, safety, livelihoods improved? |
-| **Social Cohesion Impact** | 15% | Communities strengthened, solidarity built? |
-| **Justice & Rights Impact** | 10% | Wrongs addressed, rights expanded? |
-
-### Assessment Dimensions (HOW real/accessible)
-| Dimension | Weight | Question |
-|-----------|--------|----------|
-| **Evidence Level** | 20% | Documented outcomes or speculation? |
-| **Benefit Distribution** | 20% | Who benefits? Elite → Universal? |
-| **Change Durability** | 10% | Temporary relief → Systemic change? |
-
-## Performance
-
-| Metric | Value |
-|--------|-------|
-| **Validation MAE** | **0.681** |
-| Training MAE | 0.637 |
-| Validation RMSE | 0.880 |
-
-### Per-Dimension MAE (Validation)
-| Dimension | MAE |
-|-----------|-----|
-| Human Wellbeing Impact | 0.686 |
-| Social Cohesion Impact | 0.704 |
-| Justice Rights Impact | 0.619 |
-| Evidence Level | 0.636 |
-| Benefit Distribution | 0.792 |
-| Change Durability | 0.648 |
-
-## Training Details
-
-- **Base Model**: Qwen/Qwen2.5-1.5B
-- **Training Mode**: Knowledge Distillation (from Gemini Flash oracle)
-- **Adapter**: LoRA (18.5M trainable params, 1.2% of model)
-- **Training Samples**: 7,999
-- **Validation Samples**: 1,000
-- **Epochs**: 3
-- **Batch Size**: 8
-- **Learning Rate**: 2e-5
-- **Max Length**: 512 tokens
-
-## Usage
-
-```python
-from transformers import AutoTokenizer, AutoModelForSequenceClassification
-from peft import PeftModel
-import torch
-
-# Load base model and LoRA adapter
-base_model = AutoModelForSequenceClassification.from_pretrained(
-    "Qwen/Qwen2.5-1.5B",
-    num_labels=6,
-    problem_type="regression"
-)
-model = PeftModel.from_pretrained(base_model, "nexusmind/uplifting-filter-v5")
-tokenizer = AutoTokenizer.from_pretrained("nexusmind/uplifting-filter-v5")
-
-# Score an article
-article = "Title: Community garden feeds 500 families\n\nA new community garden..."
-inputs = tokenizer(article, return_tensors="pt", max_length=512, truncation=True)
-
-with torch.no_grad():
-    outputs = model(**inputs)
-    scores = outputs.logits[0].numpy()
-
-dimensions = ["human_wellbeing_impact", "social_cohesion_impact", "justice_rights_impact",
-              "evidence_level", "benefit_distribution", "change_durability"]
-
-for dim, score in zip(dimensions, scores):
-    print(f"{dim}: {score:.1f}")
-```
-
-## Gatekeeper Rule
-
-**Evidence Level < 3 → Overall score capped at 3.0**
-
-Speculation without documented outcomes cannot be truly uplifting.
-
-## Limitations
-
-- Trained on multilingual news articles (61% English, 31% French, 7% Spanish, <1% German/Dutch/Italian)
-- MAE of ~0.68 means predictions within ±0.7 of oracle on average
-- `benefit_distribution` dimension has highest error (0.79 MAE)
-- Model focuses on documented outcomes, not emotional tone
-
-## License
-
-MIT
-
-## Citation
-
-```bibtex
-@misc{uplifting_filter_v5,
-  title={Uplifting Content Filter v5},
-  author={NexusMind},
-  year={2025},
-  url={https://huggingface.co/nexusmind/uplifting-filter-v5}
-}
-```
+
+# Uplifting Content Filter v5
+
+## Model Description
+
+A fine-tuned **Qwen2.5-1.5B** model with LoRA adapters for multi-dimensional uplifting content scoring.
+
+This model evaluates news articles across **6 orthogonal dimensions** to identify genuinely uplifting content with documented positive outcomes - not just feel-good stories or speculation.
+
+**Key Innovation**: Uses an orthogonal dimension framework (inspired by LCSA methodology) to avoid the high correlation issues found in previous versions.
+
+## Dimensions
+
+The model scores articles on 6 dimensions:
+
+### Impact Domains (WHAT kind of uplift)
+| Dimension | Weight | Question |
+|-----------|--------|----------|
+| **Human Wellbeing Impact** | 25% | Health, safety, livelihoods improved? |
+| **Social Cohesion Impact** | 15% | Communities strengthened, solidarity built? |
+| **Justice & Rights Impact** | 10% | Wrongs addressed, rights expanded? |
+
+### Assessment Dimensions (HOW real/accessible)
+| Dimension | Weight | Question |
+|-----------|--------|----------|
+| **Evidence Level** | 20% | Documented outcomes or speculation? |
+| **Benefit Distribution** | 20% | Who benefits? Elite → Universal? |
+| **Change Durability** | 10% | Temporary relief → Systemic change? |
+
+## Performance
+
+| Metric | Value |
+|--------|-------|
+| **Validation MAE** | **0.681** |
+| Training MAE | 0.637 |
+| Validation RMSE | 0.880 |
+
+### Per-Dimension MAE (Validation)
+| Dimension | MAE |
+|-----------|-----|
+| Human Wellbeing Impact | 0.686 |
+| Social Cohesion Impact | 0.704 |
+| Justice Rights Impact | 0.619 |
+| Evidence Level | 0.636 |
+| Benefit Distribution | 0.792 |
+| Change Durability | 0.648 |
+
+## Training Details
+
+- **Base Model**: Qwen/Qwen2.5-1.5B
+- **Training Mode**: Knowledge Distillation (from Gemini Flash oracle)
+- **Adapter**: LoRA (18.5M trainable params, 1.2% of model)
+- **Training Samples**: 7,999
+- **Validation Samples**: 1,000
+- **Epochs**: 3
+- **Batch Size**: 8
+- **Learning Rate**: 2e-5
+- **Max Length**: 512 tokens
+
+## Usage
+
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+from peft import PeftModel
+import torch
+
+# Load base model and LoRA adapter
+base_model = AutoModelForSequenceClassification.from_pretrained(
+    "Qwen/Qwen2.5-1.5B",
+    num_labels=6,
+    problem_type="regression"
+)
+model = PeftModel.from_pretrained(base_model, "nexusmind/uplifting-filter-v5")
+tokenizer = AutoTokenizer.from_pretrained("nexusmind/uplifting-filter-v5")
+
+# Score an article
+article = "Title: Community garden feeds 500 families\n\nA new community garden..."
+inputs = tokenizer(article, return_tensors="pt", max_length=512, truncation=True)
+
+with torch.no_grad():
+    outputs = model(**inputs)
+    scores = outputs.logits[0].numpy()
+
+dimensions = ["human_wellbeing_impact", "social_cohesion_impact", "justice_rights_impact",
+              "evidence_level", "benefit_distribution", "change_durability"]
+
+for dim, score in zip(dimensions, scores):
+    print(f"{dim}: {score:.1f}")
+```
+
+## Gatekeeper Rule
+
+**Evidence Level < 3 → Overall score capped at 3.0**
+
+Speculation without documented outcomes cannot be truly uplifting.
+
+## Limitations
+
+- Trained on multilingual news articles (61% English, 31% French, 7% Spanish, <1% German/Dutch/Italian)
+- MAE of ~0.68 means predictions within ±0.7 of oracle on average
+- `benefit_distribution` dimension has highest error (0.79 MAE)
+- Model focuses on documented outcomes, not emotional tone
+
+## License
+
+MIT
+
+## Citation
+
+```bibtex
+@misc{uplifting_filter_v5,
+  title={Uplifting Content Filter v5},
+  author={NexusMind},
+  year={2025},
+  url={https://huggingface.co/nexusmind/uplifting-filter-v5}
+}
+```
+### Framework versions
+
+- PEFT 0.17.1

adapter_config.json

@@ -19,6 +19,8 @@
   "megatron_config": null,
   "megatron_core": "megatron.core",
   "modules_to_save": [
+    "classifier",
+    "score",
     "classifier",
     "score"
   ],
@@ -28,13 +30,13 @@
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
-    "o_proj",
-    "down_proj",
     "k_proj",
     "q_proj",
+    "down_proj",
     "v_proj",
     "up_proj",
-    "gate_proj"
+    "gate_proj",
+    "o_proj"
   ],
   "target_parameters": null,
   "task_type": "SEQ_CLS",

tokenizer.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:85acc0ed1a93f8b0e6c803b53edf0fe4898ac19a3fff657f21020d280364a0cf
-size 11422174
+oid sha256:9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa
+size 11421896